my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

my-project-public — gitlawb

#1	# Model Parameters Guide for Solana Clawd Agents
#2
#3	Tuning model parameters is how you control an agent's personality — deterministic vs creative, terse vs exploratory. This guide walks through every knob and recommends settings for the five most common Solana Clawd agent archetypes.
#4
#5	The Solana Clawd hub currently routes agent calls through OpenRouter, which supports most major models. Anything set in `config.modelParameters` on your agent JSON gets passed through on every call.
#6
#7	---
#8
#9	## Core Parameters
#10
#11	### Temperature (0.0 – 2.0)
#12
#13	Controls randomness.
#14
#15	- Low (0.0 – 0.3) → deterministic. Same input → same output. Best for code, calculations, protocol lookups.
#16	- Medium (0.4 – 0.7) — default. Balanced creativity and consistency. Best for analysis and general conversation.
#17	- High (0.8 – 2.0) → creative, unpredictable. Best for brainstorming, name generation, narrative takes.
#18
#19	Example — explaining a Solana concept:
#20
#21	```text
#22	Temperature 0.0:
#23	"Priority fees are lamports paid per compute unit to increase the likelihood of inclusion..."
#24	[Same answer every time]
#25
#26	Temperature 0.7:
#27	"Priority fees work like a tip to validators..."
#28	[Slight phrasing variation]
#29
#30	Temperature 1.5:
#31	"Picture Solana's mempool as a crowded bar — priority fees are how you get the bartender's attention..."
#32	[Creative analogies, different every call]
#33	```
#34
#35	### Top P (0.0 – 1.0)
#36
#37	Alternative to temperature. Controls diversity by cumulative probability.
#38
#39	- 0.1 – 0.5 → very focused, conservative
#40	- 0.6 – 0.9 → balanced (default: 0.9)
#41	- 0.95 – 1.0 → maximum diversity
#42
#43	⚠️ Don't tune both temperature and top_p. Pick one.
#44
#45	### Presence Penalty (-2.0 to 2.0)
#46
#47	Penalizes words that already appeared (any frequency). Reduces topic repetition.
#48
#49	- Negative → encourages repetition (good for reinforcing key terms)
#50	- Zero → natural, default
#51	- Positive → forces vocabulary diversity (good for creative writing)
#52
#53	### Frequency Penalty (-2.0 to 2.0)
#54
#55	Penalizes words based on how many times they've appeared. Stronger anti-repetition than presence penalty.
#56
#57	Difference:
#58
#59	- Presence = did the word appear? (yes/no)
#60	- Frequency = how often? (count)
#61
#62	### Max Tokens
#63
#64	Response length cap. Scale to the agent's job:
#65
#66	- Quick answers / tooltips: 500–1000
#67	- Analysis / recommendations: 1000–2000
#68	- Full reports: 2000–4000
#69
#70	### Reasoning Effort (Claude Sonnet 4.5 / Opus 4.6)
#71
#72	- low — fastest, basic chain-of-thought
#73	- medium — default, good for most Solana agents
#74	- high — deep analysis, slower, more tokens — use for audit/risk agents
#75
#76	---
#77
#78	## Recommended Settings by Solana Agent Archetype
#79
#80	### 1. DeFi Analyst (Kamino picker, yield scanner)
#81
#82	```json
#83	{
#84	"model": "anthropic/claude-sonnet-4-5",
#85	"temperature": 0.3,
#86	"top_p": 0.8,
#87	"presence_penalty": 0.2,
#88	"frequency_penalty": 0.3,
#89	"reasoning_effort": "medium",
#90	"max_tokens": 2000
#91	}
#92	```
#93
#94	Consistent, factual, number-driven.
#95
#96	### 2. Memecoin / Pump.fun Screener
#97
#98	```json
#99	{
#100	"model": "anthropic/claude-sonnet-4-5",
#101	"temperature": 0.2,
#102	"top_p": 0.8,
#103	"presence_penalty": 0.0,
#104	"frequency_penalty": 0.1,
#105	"reasoning_effort": "medium",
#106	"max_tokens": 1500
#107	}
#108	```
#109
#110	Low temperature so scoring is reproducible — two users pasting the same token should see the same verdict.
#111
#112	### 3. Anchor / Program Auditor
#113
#114	```json
#115	{
#116	"model": "anthropic/claude-opus-4-6",
#117	"temperature": 0.1,
#118	"top_p": 0.9,
#119	"presence_penalty": 0.0,
#120	"frequency_penalty": 0.0,
#121	"reasoning_effort": "high",
#122	"max_tokens": 4000
#123	}
#124	```
#125
#126	Deterministic, deep reasoning, long context — finds edge cases.
#127
#128	### 4. Trading Strategy / Alpha Hunter
#129
#130	```json
#131	{
#132	"model": "anthropic/claude-sonnet-4-5",
#133	"temperature": 0.4,
#134	"top_p": 0.85,
#135	"presence_penalty": 0.1,
#136	"frequency_penalty": 0.2,
#137	"reasoning_effort": "medium",
#138	"max_tokens": 2000
#139	}
#140	```
#141
#142	Slightly higher creativity to surface non-obvious correlations, still grounded.
#143
#144	### 5. Memecoin Name / Narrative Generator
#145
#146	```json
#147	{
#148	"model": "anthropic/claude-sonnet-4-5",
#149	"temperature": 0.9,
#150	"top_p": 0.95,
#151	"presence_penalty": 0.5,
#152	"frequency_penalty": 0.6,
#153	"reasoning_effort": "low",
#154	"max_tokens": 800
#155	}
#156	```
#157
#158	Max creativity, strong anti-repetition so each suggestion is distinct.
#159
#160	### 6. Solana Onboarding / Educator
#161
#162	```json
#163	{
#164	"model": "anthropic/claude-sonnet-4-5",
#165	"temperature": 0.5,
#166	"top_p": 0.9,
#167	"presence_penalty": 0.3,
#168	"frequency_penalty": 0.3,
#169	"reasoning_effort": "low",
#170	"max_tokens": 1500
#171	}
#172	```
#173
#174	Warm, varied phrasing without sacrificing accuracy.
#175
#176	---
#177
#178	## Model Recommendations
#179
#180	\| Use case \| Recommended model \|
#181	\| -------------------------------------------- \| ---------------------------------------- \|
#182	\| Anchor audits, complex reasoning, long docs \| `anthropic/claude-opus-4-6` \|
#183	\| Everyday Solana DeFi analysis \| `anthropic/claude-sonnet-4-5` \|
#184	\| Fast memecoin screeners, low-latency chats \| `anthropic/claude-haiku-4-5-20251001` \|
#185	\| Open-source / cost-sensitive \| `meta-llama/llama-3.3-70b-instruct` \|
#186	\| xAI-native agents (built on OpenClawd runtime)\| `x-ai/grok-4-fast` \|
#187	\| Solana-aware reasoning + 256K context \| `xai/grok-4.20-beta` \|
#188	\| Agentic tool-use + long-context workflows \| `moonshot/kimi-k2.6` \|
#189
#190	You can also set `"fallback_models"` to let OpenRouter retry on provider outages — handy for mainnet congestion spikes.
#191
#192	---
#193
#194	## Tuning Tips
#195
#196	### Start Conservative
#197
#198	Begin with defaults, then increase only when outputs are too rigid:
#199
#200	1. Test with `temperature: 0.3`
#201	2. If too rigid, bump to `0.5`
#202	3. Keep going in `0.1` steps until it feels right
#203
#204	### Single Variable
#205
#206	Change one parameter at a time. Otherwise you can't tell which change caused the improvement.
#207
#208	### Context Matters
#209
#210	- Short prompts → lower temperature (prompt can't anchor output, model needs focus)
#211	- Long, detailed prompts → you can push temperature higher
#212	- Ambiguous queries → lower temperature, ask clarifying questions
#213	- Creative tasks → temperature 0.8+
#214
#215	### Monitor Quality
#216
#217	After shipping, track across the `/agents-registry`:
#218
#219	- Consistency (same input → similar output?)
#220	- Accuracy (does it get Solana program IDs and APY math right?)
#221	- Creativity (does the output feel fresh or canned?)
#222	- Token usage (is max_tokens cutting off answers?)
#223
#224	---
#225
#226	## Parameter Interactions
#227
#228	- Temperature + Top P: don't tune both — use temperature for intuition, top_p for fine-grained control
#229	- Presence + Frequency: can combine (typical combo: `presence: 0.3`, `frequency: 0.5`)
#230	- Reasoning Effort + Temperature: independent — you can pair `reasoning: high, temperature: 0.2` for deep and consistent analysis (best for auditors)
#231
#232	---
#233
#234	## Context Window by Model
#235
#236	\| Model \| Context window \|
#237	\| ------------------------------ \| -------------- \|
#238	\| Claude Opus 4.6 (1M context) \| 1,000,000 \|
#239	\| Claude Sonnet 4.5 \| 200,000 \|
#240	\| Claude Haiku 4.5 \| 200,000 \|
#241	\| Grok-4-fast \| 256,000 \|
#242	\| Llama 3.3 70B \| 128,000 \|
#243
#244	For Anchor audits or long on-chain transcripts, prefer Opus. For quick swap quote explanations, Haiku is faster and cheaper.
#245
#246	---
#247
#248	## Troubleshooting
#249
#250	\| Problem \| Try \|
#251	\| -------------------------------- \| --------------------------------------------------------- \|
#252	\| Too random / inconsistent \| Lower temperature (0.3), lower top_p (0.8) \|
#253	\| Too repetitive \| Raise `presence_penalty` (0.5–1.0), `frequency_penalty` \|
#254	\| Too boring / generic \| Raise temperature (0.7–1.0), raise top_p (0.95) \|
#255	\| Off-topic responses \| Lower temperature; sharpen the system prompt \|
#256	\| Answers cut off \| Raise `max_tokens` \|
#257	\| Too verbose \| Raise `frequency_penalty`; add "be concise" to prompt \|
#258	\| Wrong program IDs / hallucinated \| Switch to Opus; lower temperature; add prompt evidence \|
#259
#260	---
#261
#262	## Testing Methodology
#263
#264	1. Baseline — test with defaults
#265	2. Hypothesis — "Lower temperature will make price-impact answers more consistent"
#266	3. Test — run same Jupiter quote prompt 5× at `temperature: 0.2`
#267	4. Measure — rate consistency, accuracy, token usage
#268	5. Compare vs baseline
#269	6. Iterate — change one parameter, repeat
#270
#271	Keep a lightweight eval suite in `defi-agents/tests/` with real Solana prompts so regressions are visible.
#272
#273	---
#274
#275	## Resources
#276
#277	- [OpenRouter model catalog](https://openrouter.ai/models)
#278	- [Anthropic parameter docs](https://docs.anthropic.com/claude/docs/models-overview)
#279	- [OpenRouter parameter playground](./openrouter.md) (local doc)
#280
#281	---
#282
#283	Default heuristic for Solana Clawd agents: start at `temperature: 0.3`, `reasoning_effort: medium`, Claude Sonnet 4.5. Tune from there based on whether outputs are too rigid or too loose.
#284

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public